Abstract

Automatic recognition of faces is now a reality. In our experiments we seek to establish bounds on the information content required as a function of the configuration of the machine. In general with the direct image, excellent recognition is obtained for information capacity below Harmon’s (1973) benchmark value of 500. Our main data base is gathered from twenty-five people consisting of three views (front, 30° turn, 60° turn) and ten separate exposures for each view. Typically, these are grouped into a learning set of five and a testing set of five. In the first series of experiments, an optical Fourier transform configuration is used which includes a ring-wedge photodetector interface together with neural network software. Good recognition is obtained, and we report on the training using front and side views. In the second series of experiments, frame-grabbed images are used, again with the neural network software. Tests are made of the information capacity required as a function of the image coding. Comparative error rates are presented for recognition in the following cases: with 256 grey levels or using global binarization or using an error-diffusion halftoning algorithm. Particularly good results are obtained from the error-diffusion form of the imagery.

© 1991 Optical Society of America